1. The AMD and Intel Architectures COMP311 2007 Jamie Curtis

2. Intel History 4 & 8 bit Micro’s introduced in 1971 & 1972 740kHz, 2300 transistors First x86 arrives in June, 1978 – the 8086 16bit data, 20bit address 8088 variant used in the IBM PC Identical internal to 8086 with 8bit external bus 29,000 transistors 80286 introduced in 1982 Still 16bit. Introduced memory protection and protected mode 6 – 12MHz, 134,000 transistors

3. Intel History cont. 80386 introduced in October 1985 First 32bit processor Re-worked protection allows paged virtual memory Codenamed P3, called the i386 Floating point via the 80387 co-processor 275,000 Transistors 4 GB Addressable memory

4. Intel History cont. 80486 introduced in 1989 With the arrival of the 80486DX2 in 1992 for the first time the external bus no longer runs at the CPU frequency L1 Cache now on chip Pentium introduced in 1993 Intel have one of their first big public recalls, the FDIV bug in early Pentiums Pentium MMX follows in 1997

5. Intel History cont. Pentium Pro introduced in 1995 First mainstream processor that translates instructions into RISC like microinstructions before executing them Integrated L1 & L2 cache First product based on the P6 architecture Highly optimised for 32bit code made it a poor choice for the 16bit Windows 3.11

6. Intel History cont. P6 Core continues into many designs Pentium II adds MMX (1997) L2 cache ½ speed external Pentium III adds SSE (1999) L2 cache becomes integrated again

7. Intel NetBurst Intel introduces the Pentium 4 in 2000 based on the all new NetBurst (P7) architecture. all about increasing clock speed when released > 10GHz promised To achieve the high clock speeds, a very deep pipeline is required 20 stages originally, 31 stages in the Prescott core Avoiding stalls requires the “Rapid-Execution-Engine”  ALU runs at twice the core frequency Branch prediction becomes important  > 80% correctly predicted by the P4

8. Intel NetBurst Deep, wide CPU has problem keeping it’s ALU’s busy “According to Intel, most IA-32 x86 code uses only 35% of the Pentium 4's execution units” Hyper-Threading allows one CPU to act like SMP system

9. Intel NetBurst Works well for traditional “Enterprise” applications that can be parallelised well Trouble is, most code is very similar Causes collisions on execution units Some code actually runs slower NetBurst does allow execution units to be added relatively easily Fits Hyper-Threading

10. Intel NetBurst Well, what went wrong ? HEAT ! While you can scale frequency with a deeper pipeline, the heat dramatically rises with this rise in frequency. Increasing the supporting logic to allow the pipeline to work effectively also increases transistor count, all creating more heat. Prescotts contain over 125 million transistors

11. Intel Core Microarchitecture Pentium M resurrects P6 architecture again in 2003 Adds SIMD Adds NetBurst (P7) FSB Further development from the P6 leads to the Core architecture. Designed from the ground up, but the design follows the P6 much more than the NetBurst

12. Intel Core cont. Much wider design than previously Pipeline down from 30+ to 14 stages Up to 5 x86 instructions can be decoded per cycle NetBurst could only decode one instruction Massive increase in re-ordering ability Also added operation fusing L2 cache up to 4Mb 4 smarter prefetch units help keep this utilised

13. AMD History From 1979 – 1991 AMD was second sourcing Intel processors A requirement to supply to IBM Intel attempted to stop this for the 386, so AMD cloned their own version, the Am386 AMD introduce the K5 in 1995 compete with the Pentium ultimately it becomes a failure

14. AMD History cont. The first of 3 K6 variants is introduced in 1997 Backward compatible with Intel Pentium motherboards K6-III introduces on chip full speed cache, topping Intel’s ½ speed external L2 cache in the PII K6-III production is halted due to the demand for the new K7 – Athlon CPU.

15. AMD History cont. Introduced in 1999, the K7 core was renamed to the Athlon First time AMD required an incompatible motherboard from Intel, although the Athlon was introduced using a CPU SEC cartridge that was mechanically identical to Intel’s P-II cartridge.

16. AMD History cont. AMD have used “PR” ratings for their CPU’s since the K5. “Performance Ratings” are a way to combat the higher frequency of the Intel products vs the higher IPC of the AMD products. Re-introduced for the Athlon XP because of the much higher P4 clockrate.

17. AMD K8 First released as the “Opteron” for the server market in 2003 and later as the “Athlon 64” for the desktop market. First 64bit CPU that could run 32bit x86 code without a performance hit The K8 is the first x86 CPU to bring the memory controller onto the CPU die Much lower latency Less dependence on chipsets Runs at core speed

18. AMD K8 Based heavily off of the K7 design Better branch prediction allows a slightly deeper pipeline ( therefore higher clocks) 10 stages in K7, 12 in K8 Increased TLBs Allows better cache performance for large memory New FSB system HyperTransport Used to also allow much more scalable SMP

19. AMD 10h Family (K10) Plug in replacements for K8 AM2+ for Desktop, Socket F+ for Servers Multi core design Additional shared L3, individual L2 and L1 (split I & D) 128 bit FPU units Significant improvement of FP over K8 Similar enhancements to fetch, decode and prediction logic to the Intel Core

20. AMD 10h Family cont. Known as “Phenom” for the desktop, Opteron for servers First 10h Opteron was codenamed Barcelona 463M Transistors

21. AMD64 + EM64T Designed by AMD Intel focused 64 bit development on Itainum (IA- 64). Server focused EM64T reverse engineered when AMD64 became popular for entry level servers and desktops and added to P4 First major extension to the x86 ISA since i386

22. x86-64 i386 has 2 major modes Real mode (8086 emulation) Protected mode (32bit) x86-64 bundles the above into Legacy Mode Legacy Mode works with all existing code Adds Long Mode Split into full “64 bit mode” and 32bit “Compatibility Mode” Requires OS support Processes running in compatibility mode require no changes

23. x86-64 Enhancements REGISTERS !

24. x86-64 Enhancements 64bit addressing allows memory accesses above 4G without nasty hacks. “NX” (No Execute) bit on a per page basis. Improved support for Position Independent Code (by IP relative addressing modes). Adds some extra opcodes to improve common operations. Adds virtualisation features. Individual processes can be 32 or 64bit